What is Feature Importance in a Trading Model?

When you train a machine learning trading model, you give it a menu of inputs — technical indicators, fundamental data, macro signals. The model then decides, from historical data, which of those inputs actually mattered for predicting the outcome you defined.

Feature importance is how you see what it decided.

What Feature Importance Actually Shows

Feature importance is a chart that shows each input signal and how much weight the model assigned to it during training.

A feature with 15% importance means the model relied on that signal heavily when making predictions. A feature with 0.3% importance means the model essentially ignored it — it might have been included in your input menu, but the data didn't support it as a useful predictor.

This is information you can't get from a traditional rule-based trading system. When you hard-code "buy when the 50-day crosses the 100-day," there's nothing to inspect — you wrote the rule, you know what it does. With a trained model, the rules emerged from the data. Feature importance is how you read those rules.

A Real Example: Two Models, One Universe

Here's what feature importance revealed when I trained two iterations of the same model on the QB500 universe, with the same target (+2.5% in 4 days), with slightly different feature menus.

Model 1 landed on the 2-10 Year Yield Spread as its top feature at 10% weight. The rest of its weights were distributed fairly evenly across 40+ signals. No single feature dominated — it was running a committee vote every day.

Model 3 landed on QB1000 P/E Median (16.6%), Market Trend Strength (14.6%), and Nasdaq-100 RSI (11.6%) as its top three features. Three signals doing most of the heavy lifting, all of them about whether the broad market is trending and whether valuations support it.

Same universe. Same target. Slightly different inputs. Completely different worldviews.

Model 1 became a consistent daily participant — 10–20 picks every day regardless of market conditions. Model 3 became a conviction trader — barely picking during market stress, then firing 100+ picks in a single session when its trend and valuation signals all confirmed at once.

Without the feature importance chart, both models would have been equally opaque. You'd see that one is steady and one is explosive, but you wouldn't know why. With the chart, the reason is visible.

Why It Changes How You Iterate

Feature importance is what makes model iteration intelligent rather than random.

If a model isn't performing the way you expected, the feature importance chart often tells you why. Maybe one signal is dominating with 20%+ weight — your model is essentially a single-indicator strategy with extra steps. Maybe the signals you thought would matter (e.g., RSI) have near-zero importance — the model isn't using them. Maybe the top features are all macro signals when you wanted a stock-picker, or vice versa.

Each of those observations gives you a specific hypothesis to test on the next iteration: add more individual stock features, remove the over-weighted one, try a different technical indicator, constrain the macro weight. You're not adjusting randomly. You're responding to what the model is telling you about what it found.

This is the feedback loop that rule-based platforms never give you.

What to Look for When You Review Your Feature Importance

A few things worth checking after every training run:

• Concentration: Is one feature doing 20%+ of the work? That's a red flag — your model may be a single-signal strategy dressed as ML. Diversified weights across multiple features usually generalize better.
• Balance between stock and market signals: A model that's 80% macro signals will behave very differently from one that's 80% individual stock signals. Neither is wrong — but you should know which one you built.
• Surprising top features: If signals you included "just in case" ended up at the top, that's worth investigating. Sometimes the model found something you didn't expect. Sometimes it's overfitting to a quirk in the training data.
• Features with near-zero importance: These can usually be removed. A cleaner feature set often trains better models.

Feature importance doesn't tell you everything about why a model works. But it tells you more than almost any other tool available to retail traders — and it's the first thing to look at every time you train.